Spaces:
Sleeping
Sleeping
Commit History
-
feat: Aether engine speed dustoff -- PyTorch vs WASM-SIMD
TaylorSame model (SmolLM2-360M-Instruct Q8_0). Same prompt. Same tokens. Pure engine comparison. No training differences, no decoder tricks. PyTorch: ~2.8GB runtime, C++/CUDA/MKL, Python bindings Aether: 14KB WASM binary + JS, SIMD128, zero ML dependencies Both run in parallel. Whichever finishes first shows first. Cyan accent.699a27f